What the US Copyright Office Actually Requires for AI Work

There's a lot of noise about AI and copyright. Most of it is wrong, or at least imprecise enough to get you in trouble. So let's strip it down to what the United States Copyright Office actually requires, in plain terms, with the cases that back it up.

The good news is that the rule is more knowable than the headlines suggest. The bad news is that it's also stricter than most people building with AI would like, and the part that protects you is the part nobody documents.

The one requirement everything hangs on

Copyright protects works of human authorship. That's the foundation, and it isn't new. What's new is AI forcing the question of where human authorship begins and a machine's output ends.

The Office's position, set out in its 2025 guidance, is that a work containing AI-generated material can still be registered, but only the human-authored parts are protected. The purely machine-generated material is not. A human has to have authored the expressive elements you're claiming. No human author for a given element, no copyright in that element. Simple to state, harder to apply, and absolutely central to everything below.

This got tested in court. In Thaler v. Perlmutter, decided in 2025, the applicant tried to register an image with his AI system listed as the author. The court refused. Authorship requires a human. The work was left unprotectable, which is to say anyone can copy it freely. That's the cost of getting this wrong: not a smaller copyright, but no copyright at all.

Control over expressive elements, three ways

So what counts as human authorship when a machine did some of the work? The Office points to control over the expressive elements of the output, and it names three forms that control can take.

Selection. You choose which outputs to keep and which to discard. Out of fifty generations, you pick the one that works, for reasons rooted in your own judgment.

Arrangement. You combine, sequence, or compose elements into something the machine didn't produce on its own. You lay out the panels, structure the document, assemble the pieces.

Modification. You change what the machine produced. You edit, rewrite, paint over, restructure, until the expression carries your hand.

Do enough of any of these and the human contribution becomes protectable. Notice the common thread: in each case you, the human, are controlling how the work is expressed, not just what it's about. That distinction is the whole game.

Prompts alone don't count, no matter how long

This is the requirement people fight hardest, so be clear-eyed about it. A prompt, on its own, is not authorship. And the length doesn't save it.

The Office's reasoning is that when you write a prompt, you're describing a result, but the model decides the specific expression. You don't control which of countless plausible outputs it generates. A 600-word prompt full of detail about lighting, mood, and composition is still just a more elaborate request. The expressive choices, where the shadow lands, how the sentence is phrased, what the texture looks like, are made by the system, not by you.

People hate this because they put real craft into prompting. Fair. But prompt craft and copyright authorship are different things, and the law only rewards the second one.

There's no magic number

Here's a trap worth naming. People want a threshold. "If I'm 60% of the work, I'm covered." It doesn't exist.

The Office evaluates case by case. There's no percentage, no edit count, no ratio of human-to-machine input that flips protection on. Whether a work has enough human authorship is a qualitative judgment about the expressive elements a human actually controlled, made on the specific facts of that work. Any product or service that hands you a "human contribution score" and implies it determines copyrightability is inventing a standard the Office doesn't use. Be suspicious of clean numbers in a domain the law deliberately keeps qualitative.

What Zarya of the Dawn teaches

The best worked example is Zarya of the Dawn, a comic Kris Kashtanova created using Midjourney. The Office didn't give a yes-or-no verdict on the whole thing. It split it.

The text Kashtanova wrote? Protected. The selection and arrangement of the images into a coherent comic? Protected, that's human authorship. The individual images Midjourney generated? Not protected, because the human didn't author the expression in them. One registration, a clean internal line between the human-authored and machine-authored parts.

That's the model for how the Office thinks. Not "is this an AI work, yes or no," but "which specific elements did a human author, and which did the machine produce." You can own the arrangement and not the raw images. You can own the words and not the pictures. Granular, element by element.

You have to disclose and disclaim

One more requirement that trips people up at registration time. If your work contains AI-generated material, you must tell the Office. You disclose that AI was used, and you disclaim the portions you didn't author as a human.

You cannot register an AI-assisted work as though a person made every part of it by hand. Doing so risks the registration itself, because a registration built on a misstatement about authorship is vulnerable. The honest path is also the safe one: disclose the AI, claim the human authorship clearly, disclaim the rest. To do that well, you need to actually know which parts are which, which brings us to the problem hiding inside all of this.

The requirement behind the requirements

Every rule above assumes you can answer one question: what did the human author? Selection, arrangement, modification, disclosure, disclaiming, all of it depends on being able to point at specific expressive choices and say "that one was mine."

And almost nobody can, after the fact. The finished work doesn't carry that information. A document, an image, a codebase: each is just the end state. It doesn't show the fifty discarded generations, the restructuring, the rewrite, the reasoning behind the keep-this-kill-that decisions. The evidence of your authorship evaporates the moment you stop working, and the burden of proving human authorship sits with you.

This is exactly the gap Copyrightable is built to close. It captures the creative process as it happens, the selections, the arrangements, the modifications, and ties that trail to the finished Work as evidence. It scores against a published Authorship Methodology, currently v0.1, that measures the kind of control you exercised rather than pretending to compute a percentage the Office would never recognize. It's evidence, not legal advice. It won't make machine output protectable. What it does is make sure that when the Office, or a court, or an acquirer asks the one question everything depends on, you can answer it precisely instead of guessing.

The Copyright Office's requirements aren't mysterious. Human authorship, control over expressive elements through selection, arrangement, or modification, no magic threshold, full disclosure of AI use. The hard part isn't understanding them. It's being able to prove you met them, on a specific work, long after you've forgotten the details. Keep the record while you still have it.